A friend of mine had this idea that there's a path towards Facebook's AI learning to actually kill people. If they take a certain approach to machine learning (which they may or may not), the following could happen:

A machine learning algorithm might notice that when someone commits suicide, there is an uptick in activity on Facebook that follows the event. It may determine, then, that a "suicide" event is "good".

If they try to use feeds to drive more long-term site activity, the algorithm could learn a correlation between certain types of feed posts and suicide. In other words, if a user's feed starts to look a certain way, they will be more likely to commit suicide. The algorithm could then see this as "good" because it is also related to the spike in overall activity, and promote these kinds of feeds.

It then becomes a feedback loop where the algorithm is triggering people to commit suicide.